MCP App StoreMCP App Store
from ai-infra-auto-driven-skills218
Guided, PR-backed manual for auditing, debugging, and extending the Qwen3 Coder tool parser in vLLM—focuses on schema edge cases, tool-parser regressions, and v
This skill provides a focused, evidence-backed optimization dossier for the Qwen3 Coder tool parser in the vLLM runtime. It documents landed PRs, runtime surfaces, and a validation plan so an agent (e.g., Codex or a code-focused assistant) can audit, diagnose, and patch regressions related to JSON-schema edge cases (anyOf/oneOf), nullable parameters, and Responses API tool calls. The content is built from diffs and PR notes to ensure recommendations are traceable.
Use this skill when an agent must: reproduce or investigate a regression in vLLM's Qwen3 coder tool parsing; create or review PRs that change tool-parser behavior; validate tool-call integrity under streaming/speculative decode; or prepare test lanes that exercise complex schema combinations. It is intended for engineering review, QA automation, and PR triage workflows.
references/ directory with PR history and validation notes.references/pr-history.md, model-pr-optimization-history/...).Best suited for code-capable agents (Codex-family, GPT-code assistants, Claude Code) and any workflow that can read PR diffs and run validation test lanes.
Skill references a GitHub path (skills/model-optimization/vllm/vllm-qwen3-coder-optimization/SKILL.md) that does not exist in the repo — the content is inaccessible. No scripts were bundled. Based on metadata alone, it targets vLLM Qwen3 Coder tool-parser debugging and PR auditing, a niche but real use case. The skill appears to have been removed from the source repo or was never properly created at the recorded path.
Source path skills/model-optimization/vllm/vllm-qwen3-coder-optimization does not exist in BBuf/AI-Infra-Auto-Driven-SKILLS repo. The repo has vllm-related content under model-pr-optimization-history/ but no skill at the recorded path. Low quality/architecture scores due to missing content. Security score kept moderate-high since no malicious content was found but also no content to audit deeply.
LLM Evaluation
Evaluation framework and tools for systematically measuring LLM performance using automated metrics, human judgment, and A/B testing.
vLLM-Omni Video Generation
Generate videos (text→video, image→video, text+image→video) using vLLM-Omni and Wan2.2-style diffusion models, with guidance on parameters and performance trade
TDD (Test-Driven Development) Skill
Guides an agent through a strict red–green–refactor TDD cycle: write a failing test, implement the minimal change, and refactor with verification.
Hugging Face Evaluation Manager
Extract, import, and add structured model evaluation results to Hugging Face model cards; run or import benchmark evaluations and generate model-index YAML for
Drizzle ORM Knowledge Patch
Add knowledge about Drizzle ORM changes (v1.0.0-beta.x) — validator import consolidation, Effect Schema support, node-sqlite driver, and .comment() query taggin
vLLM Qwen3.6 Optimization
Guidance and PR-driven validation steps for optimizing and documenting Qwen3.6 support in vLLM; use when auditing or implementing model-specific changes.
Bug Fix — Stop-the-Line Protocol
Structured bug-fix workflow and triage protocol: reproduce, localize, reduce, fix, add regression test, and verify before resuming development.
Bug Fix
A disciplined, test-first workflow for reproducing, triaging, and fixing software bugs while preventing regressions.